Machine learning transforms how microstates are sampled

Summary

Atomistic simulations of complex molecular systems can provide key microscopic insights not easily accessible to experiments, such as folding of proteins (see the first figure), binding of small-molecule drug candidates or peptide therapeutics to selected targets, and formation of different polymorphs of molecular crystals, provided that two major hurdles are overcome. First, an accurate description of the interatomic interactions is needed, which is captured in a single potential energy function U(x), where x denotes the full set of atomic coordinates. Second, given U(x), predicting thermodynamic and other equilibrium properties of the system or estimating kinetics requires sampling a statistically sufficient number of realizations of x from the so-called Boltzmann probability distribution P(x), which is proportional to exp[−U(x)/k_BT], where T is the system temperature and k_B is Boltzmann’s constant. On page 1001 of this issue, Noé et al. (1) introduce a machine learning–based approach to address the latter of these two challenges.